What If Your Job Is to Design the Loop, Not Run It?

There is a moment, usually late, when you look up and see a screen full of half-finished AI chats. You asked the model to draft the proposal. It did, in seconds. Then you read it, fixed the tone, fed it the next instruction, pasted the result into the document, and opened a new chat to start again. The tool was fast. The evening still went. You are using AI every single day, and somehow you are still the slowest part of your own process. Here is the cause, plainly: you are sitting inside the loop, running every turn by hand. The shift worth making is to step out of it and design the loop instead.

The people getting the most from AI stopped treating it as a faster keyboard. Boris Cherny, who built Claude Code (the coding tool from Anthropic, the company behind the Claude models), put it bluntly: he does not prompt the model anymore, he writes loops that prompt it for him. His job became designing the system that runs itself. That sounds like an engineer's concern. It is a leadership one, and it lands on your desk next.

Why does AI save the task but not the day?

Because a faster task inside an unchanged process hands the time straight back to the next task. You feel it as a strange treadmill: every individual thing is quicker, and the day is just as full. The reason is that you are still the engine. You start each turn, judge each output, and decide the next move. The model waits for you, which means your attention is the thing the whole system runs on. Speed up one turn and you have simply sped up one turn.

The change underway is bigger than speed. Gartner (a research and advisory firm) expects that by 2028 around 15% of day-to-day work decisions will be made autonomously by agentic AI, software that takes a sequence of actions toward a goal rather than answering a single question, up from effectively none in 2024. The shape of the work is moving, not only its pace. The real question for a leader is which side of that loop you intend to stand on.

What is an agentic loop, really?

It is a cycle that runs without you in the middle of it. Something triggers it. It reads where things stand, decides the next useful move, does the work, checks the result, and either carries on or raises its hand for a person. Picture the difference between driving a car and running a fleet. Turn-by-turn prompting is you behind the wheel for every journey. A loop is you running the depot: you decide where the vehicles go and what good delivery looks like, and the driving happens without your hands on each wheel. I have leaned on that fleet picture for AI before, and it holds here more literally than ever.

This is the concrete mechanism behind a phrase you will have heard. Phase One, the Age of Effort: work hard, get a little more, linear growth. Phase Two, the Age of Scale: build once, sell to millions, exponential growth. Phase Three, the Age of Acceleration: output decoupled from human effort almost entirely, the phase AI unlocks. The agentic loop is what Phase Three looks like on a Tuesday. Output keeps arriving while you are in a meeting, or asleep.

What does a loop need to run safely?

A loop left to run with no structure is how money quietly disappears. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, mostly for unclear value and missing controls rather than weak models. A loop you can trust has five parts, and a missing one is usually what breaks it.

1. A trigger. Something other than you starts each run: a schedule, an event, a new message in an inbox. The moment the start depends on your attention, you are back in the loop by hand.
2. A memory it keeps. A written record the system reads at the start of every run and updates at the end. The model forgets between runs; the record remembers, so the loop builds on yesterday instead of repeating its first step forever.
3. A maker and a separate checker. The part that does the work cannot be the part that grades it. A second pass, ideally a test or a rule that cannot be talked round, decides whether the work is good enough to pass on.
4. A ceiling. A hard limit on tries, time, or spend, set before you start. A loop chasing an impossible goal with no ceiling will run all night and bill you for it.
5. A clear intent at the top. The instruction the whole loop serves, written precisely enough that the checker can tell finished from nearly-finished.

Notice that two of those five, the checker and the ceiling, exist only to keep you safe while you are not watching. That is not overhead. That is the price of stepping out of the loop with your name still on the result.

What parts of your job remain when the loop runs itself?

Two things, and they are the two ends the automation never touches. The first is intent: saying what you actually want with enough precision that a check, human or test, can tell whether the loop delivered it. Vague intent produces a loop that runs beautifully toward the wrong place. The second is accountability: owning what ships. The loop's "done" is a claim, not a proof, and the signature on the outcome stays yours.

This is why the work points back at you before it points at the tools. The constraint was never the model's speed; it was the clarity of the person deciding what the work is for. A loop is a mirror: it scales whatever intent you give it, sharp or fuzzy, many times a day. The leaders who win the next decade are the ones who upgrade themselves first, because a clear instruction repeated a thousand times becomes an asset, and a muddled one repeated a thousand times becomes a liability.

So the practical first move is small. Pick one task you currently repeat by hand with AI: the weekly report, the inbox triage, the first draft of almost anything. Write down what good looks like, decide what checks it, set a ceiling, and design the loop around it. Run it once with you watching. Then let it run while you give your attention to the work only you can do. That is the move from operating the machine to directing it, and it starts with a single loop.

About the author

Thomas Green

British technology futurist, AI keynote speaker and advisor. Thirty years across enterprise technology and AI strategy, helping leaders navigate the future of work. The futurist who died.